DeepSeek R1 is out: a frontier reasoning model with published weights
Released today under a permissive licence, with a reported training cost far below what the field assumed was necessary. Two claims, two different levels of confidence.
DeepSeek released R1 today, 20 January 2025: a reasoning model published under an open licence, reported to approach OpenAI’s o1 on mathematics and coding benchmarks. The accompanying claim that has drawn most attention is the training cost — a figure in the single-digit millions, against the hundreds of millions assumed for comparable Western models.
Separate those two claims, because they deserve different levels of confidence.
What is verifiable
The weights are published. Anyone can download them, run them and test them. Within days there will be independent evaluations, and the performance claim will be confirmed, qualified or dismantled in public. That is how open releases work, and it is a real advantage over a press release about a closed model.
What is not
The cost figure. It describes one training run and excludes the research, the failed attempts, the infrastructure and the staff. It is not a total cost of development, and reporting it as one is a category error that will be repeated widely this week.
Treat it as directionally interesting and specifically unverified.
What an SME takes from this
Three things, none urgent.
The capability gap is now months, not years. A reasoning model of this class went from exclusive to downloadable in roughly four months. Plan on that continuing.
Your negotiating position improved. Whatever you are paying for reasoning capability, there is now a published alternative. You may never use it; it still changes the conversation at renewal.
The hosted service and the weights are different products. This matters and will be widely confused. Using DeepSeek’s own app or API means sending your text to that operator under Chinese jurisdiction. Downloading the weights and running them in a Swiss data centre is a completely different data path with completely different answers on residency.
If anyone in your company starts using the app this week — and some will, because it is free and it is about to be everywhere — that is the distinction to explain. Not « do not use it », which nobody follows, but « here is what happens to the text you type, and here is what may go in ».